Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ryusuke...
[linux-2.6/mini2440.git] / net / ipv4 / tcp_output.c
blobfcd278a7080e1656060cd5f57e38764bec3bdb74
1 /*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #include <net/tcp.h>
39 #include <linux/compiler.h>
40 #include <linux/module.h>
42 /* People can turn this off for buggy TCP's found in printers etc. */
43 int sysctl_tcp_retrans_collapse __read_mostly = 1;
45 /* People can turn this on to work with those rare, broken TCPs that
46 * interpret the window field as a signed quantity.
48 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
50 /* This limits the percentage of the congestion window which we
51 * will allow a single TSO frame to consume. Building TSO frames
52 * which are too large can cause TCP streams to be bursty.
54 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
56 int sysctl_tcp_mtu_probing __read_mostly = 0;
57 int sysctl_tcp_base_mss __read_mostly = 512;
59 /* By default, RFC2861 behavior. */
60 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
62 /* Account for new data that has been sent to the network. */
63 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
65 struct tcp_sock *tp = tcp_sk(sk);
66 unsigned int prior_packets = tp->packets_out;
68 tcp_advance_send_head(sk, skb);
69 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
71 /* Don't override Nagle indefinately with F-RTO */
72 if (tp->frto_counter == 2)
73 tp->frto_counter = 3;
75 tp->packets_out += tcp_skb_pcount(skb);
76 if (!prior_packets)
77 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
78 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
81 /* SND.NXT, if window was not shrunk.
82 * If window has been shrunk, what should we make? It is not clear at all.
83 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
84 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
85 * invalid. OK, let's make this for now:
87 static inline __u32 tcp_acceptable_seq(struct sock *sk)
89 struct tcp_sock *tp = tcp_sk(sk);
91 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
92 return tp->snd_nxt;
93 else
94 return tcp_wnd_end(tp);
97 /* Calculate mss to advertise in SYN segment.
98 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
100 * 1. It is independent of path mtu.
101 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
102 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
103 * attached devices, because some buggy hosts are confused by
104 * large MSS.
105 * 4. We do not make 3, we advertise MSS, calculated from first
106 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
107 * This may be overridden via information stored in routing table.
108 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
109 * probably even Jumbo".
111 static __u16 tcp_advertise_mss(struct sock *sk)
113 struct tcp_sock *tp = tcp_sk(sk);
114 struct dst_entry *dst = __sk_dst_get(sk);
115 int mss = tp->advmss;
117 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
118 mss = dst_metric(dst, RTAX_ADVMSS);
119 tp->advmss = mss;
122 return (__u16)mss;
125 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
126 * This is the first part of cwnd validation mechanism. */
127 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
129 struct tcp_sock *tp = tcp_sk(sk);
130 s32 delta = tcp_time_stamp - tp->lsndtime;
131 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
132 u32 cwnd = tp->snd_cwnd;
134 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
136 tp->snd_ssthresh = tcp_current_ssthresh(sk);
137 restart_cwnd = min(restart_cwnd, cwnd);
139 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
140 cwnd >>= 1;
141 tp->snd_cwnd = max(cwnd, restart_cwnd);
142 tp->snd_cwnd_stamp = tcp_time_stamp;
143 tp->snd_cwnd_used = 0;
146 /* Congestion state accounting after a packet has been sent. */
147 static void tcp_event_data_sent(struct tcp_sock *tp,
148 struct sk_buff *skb, struct sock *sk)
150 struct inet_connection_sock *icsk = inet_csk(sk);
151 const u32 now = tcp_time_stamp;
153 if (sysctl_tcp_slow_start_after_idle &&
154 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
155 tcp_cwnd_restart(sk, __sk_dst_get(sk));
157 tp->lsndtime = now;
159 /* If it is a reply for ato after last received
160 * packet, enter pingpong mode.
162 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
163 icsk->icsk_ack.pingpong = 1;
166 /* Account for an ACK we sent. */
167 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
169 tcp_dec_quickack_mode(sk, pkts);
170 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
173 /* Determine a window scaling and initial window to offer.
174 * Based on the assumption that the given amount of space
175 * will be offered. Store the results in the tp structure.
176 * NOTE: for smooth operation initial space offering should
177 * be a multiple of mss if possible. We assume here that mss >= 1.
178 * This MUST be enforced by all callers.
180 void tcp_select_initial_window(int __space, __u32 mss,
181 __u32 *rcv_wnd, __u32 *window_clamp,
182 int wscale_ok, __u8 *rcv_wscale)
184 unsigned int space = (__space < 0 ? 0 : __space);
186 /* If no clamp set the clamp to the max possible scaled window */
187 if (*window_clamp == 0)
188 (*window_clamp) = (65535 << 14);
189 space = min(*window_clamp, space);
191 /* Quantize space offering to a multiple of mss if possible. */
192 if (space > mss)
193 space = (space / mss) * mss;
195 /* NOTE: offering an initial window larger than 32767
196 * will break some buggy TCP stacks. If the admin tells us
197 * it is likely we could be speaking with such a buggy stack
198 * we will truncate our initial window offering to 32K-1
199 * unless the remote has sent us a window scaling option,
200 * which we interpret as a sign the remote TCP is not
201 * misinterpreting the window field as a signed quantity.
203 if (sysctl_tcp_workaround_signed_windows)
204 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
205 else
206 (*rcv_wnd) = space;
208 (*rcv_wscale) = 0;
209 if (wscale_ok) {
210 /* Set window scaling on max possible window
211 * See RFC1323 for an explanation of the limit to 14
213 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
214 space = min_t(u32, space, *window_clamp);
215 while (space > 65535 && (*rcv_wscale) < 14) {
216 space >>= 1;
217 (*rcv_wscale)++;
221 /* Set initial window to value enough for senders,
222 * following RFC2414. Senders, not following this RFC,
223 * will be satisfied with 2.
225 if (mss > (1 << *rcv_wscale)) {
226 int init_cwnd = 4;
227 if (mss > 1460 * 3)
228 init_cwnd = 2;
229 else if (mss > 1460)
230 init_cwnd = 3;
231 if (*rcv_wnd > init_cwnd * mss)
232 *rcv_wnd = init_cwnd * mss;
235 /* Set the clamp no higher than max representable value */
236 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
239 /* Chose a new window to advertise, update state in tcp_sock for the
240 * socket, and return result with RFC1323 scaling applied. The return
241 * value can be stuffed directly into th->window for an outgoing
242 * frame.
244 static u16 tcp_select_window(struct sock *sk)
246 struct tcp_sock *tp = tcp_sk(sk);
247 u32 cur_win = tcp_receive_window(tp);
248 u32 new_win = __tcp_select_window(sk);
250 /* Never shrink the offered window */
251 if (new_win < cur_win) {
252 /* Danger Will Robinson!
253 * Don't update rcv_wup/rcv_wnd here or else
254 * we will not be able to advertise a zero
255 * window in time. --DaveM
257 * Relax Will Robinson.
259 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
261 tp->rcv_wnd = new_win;
262 tp->rcv_wup = tp->rcv_nxt;
264 /* Make sure we do not exceed the maximum possible
265 * scaled window.
267 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
268 new_win = min(new_win, MAX_TCP_WINDOW);
269 else
270 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
272 /* RFC1323 scaling applied */
273 new_win >>= tp->rx_opt.rcv_wscale;
275 /* If we advertise zero window, disable fast path. */
276 if (new_win == 0)
277 tp->pred_flags = 0;
279 return new_win;
282 /* Packet ECN state for a SYN-ACK */
283 static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
285 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
286 if (!(tp->ecn_flags & TCP_ECN_OK))
287 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
290 /* Packet ECN state for a SYN. */
291 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
293 struct tcp_sock *tp = tcp_sk(sk);
295 tp->ecn_flags = 0;
296 if (sysctl_tcp_ecn == 1) {
297 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
298 tp->ecn_flags = TCP_ECN_OK;
302 static __inline__ void
303 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
305 if (inet_rsk(req)->ecn_ok)
306 th->ece = 1;
309 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
310 * be sent.
312 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
313 int tcp_header_len)
315 struct tcp_sock *tp = tcp_sk(sk);
317 if (tp->ecn_flags & TCP_ECN_OK) {
318 /* Not-retransmitted data segment: set ECT and inject CWR. */
319 if (skb->len != tcp_header_len &&
320 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
321 INET_ECN_xmit(sk);
322 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
323 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
324 tcp_hdr(skb)->cwr = 1;
325 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
327 } else {
328 /* ACK or retransmitted segment: clear ECT|CE */
329 INET_ECN_dontxmit(sk);
331 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
332 tcp_hdr(skb)->ece = 1;
336 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
337 * auto increment end seqno.
339 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
341 skb->csum = 0;
343 TCP_SKB_CB(skb)->flags = flags;
344 TCP_SKB_CB(skb)->sacked = 0;
346 skb_shinfo(skb)->gso_segs = 1;
347 skb_shinfo(skb)->gso_size = 0;
348 skb_shinfo(skb)->gso_type = 0;
350 TCP_SKB_CB(skb)->seq = seq;
351 if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
352 seq++;
353 TCP_SKB_CB(skb)->end_seq = seq;
356 static inline int tcp_urg_mode(const struct tcp_sock *tp)
358 return tp->snd_una != tp->snd_up;
361 #define OPTION_SACK_ADVERTISE (1 << 0)
362 #define OPTION_TS (1 << 1)
363 #define OPTION_MD5 (1 << 2)
364 #define OPTION_WSCALE (1 << 3)
366 struct tcp_out_options {
367 u8 options; /* bit field of OPTION_* */
368 u8 ws; /* window scale, 0 to disable */
369 u8 num_sack_blocks; /* number of SACK blocks to include */
370 u16 mss; /* 0 to disable */
371 __u32 tsval, tsecr; /* need to include OPTION_TS */
374 /* Write previously computed TCP options to the packet.
376 * Beware: Something in the Internet is very sensitive to the ordering of
377 * TCP options, we learned this through the hard way, so be careful here.
378 * Luckily we can at least blame others for their non-compliance but from
379 * inter-operatibility perspective it seems that we're somewhat stuck with
380 * the ordering which we have been using if we want to keep working with
381 * those broken things (not that it currently hurts anybody as there isn't
382 * particular reason why the ordering would need to be changed).
384 * At least SACK_PERM as the first option is known to lead to a disaster
385 * (but it may well be that other scenarios fail similarly).
387 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
388 const struct tcp_out_options *opts,
389 __u8 **md5_hash) {
390 if (unlikely(OPTION_MD5 & opts->options)) {
391 *ptr++ = htonl((TCPOPT_NOP << 24) |
392 (TCPOPT_NOP << 16) |
393 (TCPOPT_MD5SIG << 8) |
394 TCPOLEN_MD5SIG);
395 *md5_hash = (__u8 *)ptr;
396 ptr += 4;
397 } else {
398 *md5_hash = NULL;
401 if (unlikely(opts->mss)) {
402 *ptr++ = htonl((TCPOPT_MSS << 24) |
403 (TCPOLEN_MSS << 16) |
404 opts->mss);
407 if (likely(OPTION_TS & opts->options)) {
408 if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) {
409 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
410 (TCPOLEN_SACK_PERM << 16) |
411 (TCPOPT_TIMESTAMP << 8) |
412 TCPOLEN_TIMESTAMP);
413 } else {
414 *ptr++ = htonl((TCPOPT_NOP << 24) |
415 (TCPOPT_NOP << 16) |
416 (TCPOPT_TIMESTAMP << 8) |
417 TCPOLEN_TIMESTAMP);
419 *ptr++ = htonl(opts->tsval);
420 *ptr++ = htonl(opts->tsecr);
423 if (unlikely(OPTION_SACK_ADVERTISE & opts->options &&
424 !(OPTION_TS & opts->options))) {
425 *ptr++ = htonl((TCPOPT_NOP << 24) |
426 (TCPOPT_NOP << 16) |
427 (TCPOPT_SACK_PERM << 8) |
428 TCPOLEN_SACK_PERM);
431 if (unlikely(OPTION_WSCALE & opts->options)) {
432 *ptr++ = htonl((TCPOPT_NOP << 24) |
433 (TCPOPT_WINDOW << 16) |
434 (TCPOLEN_WINDOW << 8) |
435 opts->ws);
438 if (unlikely(opts->num_sack_blocks)) {
439 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
440 tp->duplicate_sack : tp->selective_acks;
441 int this_sack;
443 *ptr++ = htonl((TCPOPT_NOP << 24) |
444 (TCPOPT_NOP << 16) |
445 (TCPOPT_SACK << 8) |
446 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
447 TCPOLEN_SACK_PERBLOCK)));
449 for (this_sack = 0; this_sack < opts->num_sack_blocks;
450 ++this_sack) {
451 *ptr++ = htonl(sp[this_sack].start_seq);
452 *ptr++ = htonl(sp[this_sack].end_seq);
455 tp->rx_opt.dsack = 0;
459 /* Compute TCP options for SYN packets. This is not the final
460 * network wire format yet.
462 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
463 struct tcp_out_options *opts,
464 struct tcp_md5sig_key **md5) {
465 struct tcp_sock *tp = tcp_sk(sk);
466 unsigned size = 0;
468 #ifdef CONFIG_TCP_MD5SIG
469 *md5 = tp->af_specific->md5_lookup(sk, sk);
470 if (*md5) {
471 opts->options |= OPTION_MD5;
472 size += TCPOLEN_MD5SIG_ALIGNED;
474 #else
475 *md5 = NULL;
476 #endif
478 /* We always get an MSS option. The option bytes which will be seen in
479 * normal data packets should timestamps be used, must be in the MSS
480 * advertised. But we subtract them from tp->mss_cache so that
481 * calculations in tcp_sendmsg are simpler etc. So account for this
482 * fact here if necessary. If we don't do this correctly, as a
483 * receiver we won't recognize data packets as being full sized when we
484 * should, and thus we won't abide by the delayed ACK rules correctly.
485 * SACKs don't matter, we never delay an ACK when we have any of those
486 * going out. */
487 opts->mss = tcp_advertise_mss(sk);
488 size += TCPOLEN_MSS_ALIGNED;
490 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
491 opts->options |= OPTION_TS;
492 opts->tsval = TCP_SKB_CB(skb)->when;
493 opts->tsecr = tp->rx_opt.ts_recent;
494 size += TCPOLEN_TSTAMP_ALIGNED;
496 if (likely(sysctl_tcp_window_scaling)) {
497 opts->ws = tp->rx_opt.rcv_wscale;
498 opts->options |= OPTION_WSCALE;
499 size += TCPOLEN_WSCALE_ALIGNED;
501 if (likely(sysctl_tcp_sack)) {
502 opts->options |= OPTION_SACK_ADVERTISE;
503 if (unlikely(!(OPTION_TS & opts->options)))
504 size += TCPOLEN_SACKPERM_ALIGNED;
507 return size;
510 /* Set up TCP options for SYN-ACKs. */
511 static unsigned tcp_synack_options(struct sock *sk,
512 struct request_sock *req,
513 unsigned mss, struct sk_buff *skb,
514 struct tcp_out_options *opts,
515 struct tcp_md5sig_key **md5) {
516 unsigned size = 0;
517 struct inet_request_sock *ireq = inet_rsk(req);
518 char doing_ts;
520 #ifdef CONFIG_TCP_MD5SIG
521 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
522 if (*md5) {
523 opts->options |= OPTION_MD5;
524 size += TCPOLEN_MD5SIG_ALIGNED;
526 #else
527 *md5 = NULL;
528 #endif
530 /* we can't fit any SACK blocks in a packet with MD5 + TS
531 options. There was discussion about disabling SACK rather than TS in
532 order to fit in better with old, buggy kernels, but that was deemed
533 to be unnecessary. */
534 doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok);
536 opts->mss = mss;
537 size += TCPOLEN_MSS_ALIGNED;
539 if (likely(ireq->wscale_ok)) {
540 opts->ws = ireq->rcv_wscale;
541 opts->options |= OPTION_WSCALE;
542 size += TCPOLEN_WSCALE_ALIGNED;
544 if (likely(doing_ts)) {
545 opts->options |= OPTION_TS;
546 opts->tsval = TCP_SKB_CB(skb)->when;
547 opts->tsecr = req->ts_recent;
548 size += TCPOLEN_TSTAMP_ALIGNED;
550 if (likely(ireq->sack_ok)) {
551 opts->options |= OPTION_SACK_ADVERTISE;
552 if (unlikely(!doing_ts))
553 size += TCPOLEN_SACKPERM_ALIGNED;
556 return size;
559 /* Compute TCP options for ESTABLISHED sockets. This is not the
560 * final wire format yet.
562 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
563 struct tcp_out_options *opts,
564 struct tcp_md5sig_key **md5) {
565 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
566 struct tcp_sock *tp = tcp_sk(sk);
567 unsigned size = 0;
568 unsigned int eff_sacks;
570 #ifdef CONFIG_TCP_MD5SIG
571 *md5 = tp->af_specific->md5_lookup(sk, sk);
572 if (unlikely(*md5)) {
573 opts->options |= OPTION_MD5;
574 size += TCPOLEN_MD5SIG_ALIGNED;
576 #else
577 *md5 = NULL;
578 #endif
580 if (likely(tp->rx_opt.tstamp_ok)) {
581 opts->options |= OPTION_TS;
582 opts->tsval = tcb ? tcb->when : 0;
583 opts->tsecr = tp->rx_opt.ts_recent;
584 size += TCPOLEN_TSTAMP_ALIGNED;
587 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
588 if (unlikely(eff_sacks)) {
589 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
590 opts->num_sack_blocks =
591 min_t(unsigned, eff_sacks,
592 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
593 TCPOLEN_SACK_PERBLOCK);
594 size += TCPOLEN_SACK_BASE_ALIGNED +
595 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
598 return size;
601 /* This routine actually transmits TCP packets queued in by
602 * tcp_do_sendmsg(). This is used by both the initial
603 * transmission and possible later retransmissions.
604 * All SKB's seen here are completely headerless. It is our
605 * job to build the TCP header, and pass the packet down to
606 * IP so it can do the same plus pass the packet off to the
607 * device.
609 * We are working here with either a clone of the original
610 * SKB, or a fresh unique copy made by the retransmit engine.
612 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
613 gfp_t gfp_mask)
615 const struct inet_connection_sock *icsk = inet_csk(sk);
616 struct inet_sock *inet;
617 struct tcp_sock *tp;
618 struct tcp_skb_cb *tcb;
619 struct tcp_out_options opts;
620 unsigned tcp_options_size, tcp_header_size;
621 struct tcp_md5sig_key *md5;
622 __u8 *md5_hash_location;
623 struct tcphdr *th;
624 int err;
626 BUG_ON(!skb || !tcp_skb_pcount(skb));
628 /* If congestion control is doing timestamping, we must
629 * take such a timestamp before we potentially clone/copy.
631 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
632 __net_timestamp(skb);
634 if (likely(clone_it)) {
635 if (unlikely(skb_cloned(skb)))
636 skb = pskb_copy(skb, gfp_mask);
637 else
638 skb = skb_clone(skb, gfp_mask);
639 if (unlikely(!skb))
640 return -ENOBUFS;
643 inet = inet_sk(sk);
644 tp = tcp_sk(sk);
645 tcb = TCP_SKB_CB(skb);
646 memset(&opts, 0, sizeof(opts));
648 if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
649 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
650 else
651 tcp_options_size = tcp_established_options(sk, skb, &opts,
652 &md5);
653 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
655 if (tcp_packets_in_flight(tp) == 0)
656 tcp_ca_event(sk, CA_EVENT_TX_START);
658 skb_push(skb, tcp_header_size);
659 skb_reset_transport_header(skb);
660 skb_set_owner_w(skb, sk);
662 /* Build TCP header and checksum it. */
663 th = tcp_hdr(skb);
664 th->source = inet->sport;
665 th->dest = inet->dport;
666 th->seq = htonl(tcb->seq);
667 th->ack_seq = htonl(tp->rcv_nxt);
668 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
669 tcb->flags);
671 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
672 /* RFC1323: The window in SYN & SYN/ACK segments
673 * is never scaled.
675 th->window = htons(min(tp->rcv_wnd, 65535U));
676 } else {
677 th->window = htons(tcp_select_window(sk));
679 th->check = 0;
680 th->urg_ptr = 0;
682 /* The urg_mode check is necessary during a below snd_una win probe */
683 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
684 if (before(tp->snd_up, tcb->seq + 0x10000)) {
685 th->urg_ptr = htons(tp->snd_up - tcb->seq);
686 th->urg = 1;
687 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
688 th->urg_ptr = 0xFFFF;
689 th->urg = 1;
693 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
694 if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
695 TCP_ECN_send(sk, skb, tcp_header_size);
697 #ifdef CONFIG_TCP_MD5SIG
698 /* Calculate the MD5 hash, as we have all we need now */
699 if (md5) {
700 sk->sk_route_caps &= ~NETIF_F_GSO_MASK;
701 tp->af_specific->calc_md5_hash(md5_hash_location,
702 md5, sk, NULL, skb);
704 #endif
706 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
708 if (likely(tcb->flags & TCPCB_FLAG_ACK))
709 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
711 if (skb->len != tcp_header_size)
712 tcp_event_data_sent(tp, skb, sk);
714 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
715 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
717 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
718 if (likely(err <= 0))
719 return err;
721 tcp_enter_cwr(sk, 1);
723 return net_xmit_eval(err);
726 /* This routine just queues the buffer for sending.
728 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
729 * otherwise socket can stall.
731 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
733 struct tcp_sock *tp = tcp_sk(sk);
735 /* Advance write_seq and place onto the write_queue. */
736 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
737 skb_header_release(skb);
738 tcp_add_write_queue_tail(sk, skb);
739 sk->sk_wmem_queued += skb->truesize;
740 sk_mem_charge(sk, skb->truesize);
743 /* Initialize TSO segments for a packet. */
744 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
745 unsigned int mss_now)
747 if (skb->len <= mss_now || !sk_can_gso(sk) ||
748 skb->ip_summed == CHECKSUM_NONE) {
749 /* Avoid the costly divide in the normal
750 * non-TSO case.
752 skb_shinfo(skb)->gso_segs = 1;
753 skb_shinfo(skb)->gso_size = 0;
754 skb_shinfo(skb)->gso_type = 0;
755 } else {
756 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
757 skb_shinfo(skb)->gso_size = mss_now;
758 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
762 /* When a modification to fackets out becomes necessary, we need to check
763 * skb is counted to fackets_out or not.
765 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
766 int decr)
768 struct tcp_sock *tp = tcp_sk(sk);
770 if (!tp->sacked_out || tcp_is_reno(tp))
771 return;
773 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
774 tp->fackets_out -= decr;
777 /* Pcount in the middle of the write queue got changed, we need to do various
778 * tweaks to fix counters
780 static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
782 struct tcp_sock *tp = tcp_sk(sk);
784 tp->packets_out -= decr;
786 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
787 tp->sacked_out -= decr;
788 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
789 tp->retrans_out -= decr;
790 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
791 tp->lost_out -= decr;
793 /* Reno case is special. Sigh... */
794 if (tcp_is_reno(tp) && decr > 0)
795 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
797 tcp_adjust_fackets_out(sk, skb, decr);
799 if (tp->lost_skb_hint &&
800 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
801 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
802 tp->lost_cnt_hint -= decr;
804 tcp_verify_left_out(tp);
807 /* Function to create two new TCP segments. Shrinks the given segment
808 * to the specified size and appends a new segment with the rest of the
809 * packet to the list. This won't be called frequently, I hope.
810 * Remember, these are still headerless SKBs at this point.
812 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
813 unsigned int mss_now)
815 struct tcp_sock *tp = tcp_sk(sk);
816 struct sk_buff *buff;
817 int nsize, old_factor;
818 int nlen;
819 u8 flags;
821 BUG_ON(len > skb->len);
823 nsize = skb_headlen(skb) - len;
824 if (nsize < 0)
825 nsize = 0;
827 if (skb_cloned(skb) &&
828 skb_is_nonlinear(skb) &&
829 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
830 return -ENOMEM;
832 /* Get a new skb... force flag on. */
833 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
834 if (buff == NULL)
835 return -ENOMEM; /* We'll just try again later. */
837 sk->sk_wmem_queued += buff->truesize;
838 sk_mem_charge(sk, buff->truesize);
839 nlen = skb->len - len - nsize;
840 buff->truesize += nlen;
841 skb->truesize -= nlen;
843 /* Correct the sequence numbers. */
844 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
845 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
846 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
848 /* PSH and FIN should only be set in the second packet. */
849 flags = TCP_SKB_CB(skb)->flags;
850 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
851 TCP_SKB_CB(buff)->flags = flags;
852 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
854 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
855 /* Copy and checksum data tail into the new buffer. */
856 buff->csum = csum_partial_copy_nocheck(skb->data + len,
857 skb_put(buff, nsize),
858 nsize, 0);
860 skb_trim(skb, len);
862 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
863 } else {
864 skb->ip_summed = CHECKSUM_PARTIAL;
865 skb_split(skb, buff, len);
868 buff->ip_summed = skb->ip_summed;
870 /* Looks stupid, but our code really uses when of
871 * skbs, which it never sent before. --ANK
873 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
874 buff->tstamp = skb->tstamp;
876 old_factor = tcp_skb_pcount(skb);
878 /* Fix up tso_factor for both original and new SKB. */
879 tcp_set_skb_tso_segs(sk, skb, mss_now);
880 tcp_set_skb_tso_segs(sk, buff, mss_now);
882 /* If this packet has been sent out already, we must
883 * adjust the various packet counters.
885 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
886 int diff = old_factor - tcp_skb_pcount(skb) -
887 tcp_skb_pcount(buff);
889 if (diff)
890 tcp_adjust_pcount(sk, skb, diff);
893 /* Link BUFF into the send queue. */
894 skb_header_release(buff);
895 tcp_insert_write_queue_after(skb, buff, sk);
897 return 0;
900 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
901 * eventually). The difference is that pulled data not copied, but
902 * immediately discarded.
904 static void __pskb_trim_head(struct sk_buff *skb, int len)
906 int i, k, eat;
908 eat = len;
909 k = 0;
910 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
911 if (skb_shinfo(skb)->frags[i].size <= eat) {
912 put_page(skb_shinfo(skb)->frags[i].page);
913 eat -= skb_shinfo(skb)->frags[i].size;
914 } else {
915 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
916 if (eat) {
917 skb_shinfo(skb)->frags[k].page_offset += eat;
918 skb_shinfo(skb)->frags[k].size -= eat;
919 eat = 0;
921 k++;
924 skb_shinfo(skb)->nr_frags = k;
926 skb_reset_tail_pointer(skb);
927 skb->data_len -= len;
928 skb->len = skb->data_len;
931 /* Remove acked data from a packet in the transmit queue. */
932 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
934 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
935 return -ENOMEM;
937 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
938 if (unlikely(len < skb_headlen(skb)))
939 __skb_pull(skb, len);
940 else
941 __pskb_trim_head(skb, len - skb_headlen(skb));
943 TCP_SKB_CB(skb)->seq += len;
944 skb->ip_summed = CHECKSUM_PARTIAL;
946 skb->truesize -= len;
947 sk->sk_wmem_queued -= len;
948 sk_mem_uncharge(sk, len);
949 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
951 /* Any change of skb->len requires recalculation of tso
952 * factor and mss.
954 if (tcp_skb_pcount(skb) > 1)
955 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
957 return 0;
960 /* Calculate MSS. Not accounting for SACKs here. */
961 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
963 struct tcp_sock *tp = tcp_sk(sk);
964 struct inet_connection_sock *icsk = inet_csk(sk);
965 int mss_now;
967 /* Calculate base mss without TCP options:
968 It is MMS_S - sizeof(tcphdr) of rfc1122
970 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
972 /* Clamp it (mss_clamp does not include tcp options) */
973 if (mss_now > tp->rx_opt.mss_clamp)
974 mss_now = tp->rx_opt.mss_clamp;
976 /* Now subtract optional transport overhead */
977 mss_now -= icsk->icsk_ext_hdr_len;
979 /* Then reserve room for full set of TCP options and 8 bytes of data */
980 if (mss_now < 48)
981 mss_now = 48;
983 /* Now subtract TCP options size, not including SACKs */
984 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
986 return mss_now;
989 /* Inverse of above */
990 int tcp_mss_to_mtu(struct sock *sk, int mss)
992 struct tcp_sock *tp = tcp_sk(sk);
993 struct inet_connection_sock *icsk = inet_csk(sk);
994 int mtu;
996 mtu = mss +
997 tp->tcp_header_len +
998 icsk->icsk_ext_hdr_len +
999 icsk->icsk_af_ops->net_header_len;
1001 return mtu;
1004 /* MTU probing init per socket */
1005 void tcp_mtup_init(struct sock *sk)
1007 struct tcp_sock *tp = tcp_sk(sk);
1008 struct inet_connection_sock *icsk = inet_csk(sk);
1010 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1011 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1012 icsk->icsk_af_ops->net_header_len;
1013 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1014 icsk->icsk_mtup.probe_size = 0;
1017 /* This function synchronize snd mss to current pmtu/exthdr set.
1019 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1020 for TCP options, but includes only bare TCP header.
1022 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1023 It is minimum of user_mss and mss received with SYN.
1024 It also does not include TCP options.
1026 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1028 tp->mss_cache is current effective sending mss, including
1029 all tcp options except for SACKs. It is evaluated,
1030 taking into account current pmtu, but never exceeds
1031 tp->rx_opt.mss_clamp.
1033 NOTE1. rfc1122 clearly states that advertised MSS
1034 DOES NOT include either tcp or ip options.
1036 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1037 are READ ONLY outside this function. --ANK (980731)
1039 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1041 struct tcp_sock *tp = tcp_sk(sk);
1042 struct inet_connection_sock *icsk = inet_csk(sk);
1043 int mss_now;
1045 if (icsk->icsk_mtup.search_high > pmtu)
1046 icsk->icsk_mtup.search_high = pmtu;
1048 mss_now = tcp_mtu_to_mss(sk, pmtu);
1049 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1051 /* And store cached results */
1052 icsk->icsk_pmtu_cookie = pmtu;
1053 if (icsk->icsk_mtup.enabled)
1054 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1055 tp->mss_cache = mss_now;
1057 return mss_now;
1060 /* Compute the current effective MSS, taking SACKs and IP options,
1061 * and even PMTU discovery events into account.
1063 unsigned int tcp_current_mss(struct sock *sk)
1065 struct tcp_sock *tp = tcp_sk(sk);
1066 struct dst_entry *dst = __sk_dst_get(sk);
1067 u32 mss_now;
1068 unsigned header_len;
1069 struct tcp_out_options opts;
1070 struct tcp_md5sig_key *md5;
1072 mss_now = tp->mss_cache;
1074 if (dst) {
1075 u32 mtu = dst_mtu(dst);
1076 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1077 mss_now = tcp_sync_mss(sk, mtu);
1080 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1081 sizeof(struct tcphdr);
1082 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1083 * some common options. If this is an odd packet (because we have SACK
1084 * blocks etc) then our calculated header_len will be different, and
1085 * we have to adjust mss_now correspondingly */
1086 if (header_len != tp->tcp_header_len) {
1087 int delta = (int) header_len - tp->tcp_header_len;
1088 mss_now -= delta;
1091 return mss_now;
1094 /* Congestion window validation. (RFC2861) */
1095 static void tcp_cwnd_validate(struct sock *sk)
1097 struct tcp_sock *tp = tcp_sk(sk);
1099 if (tp->packets_out >= tp->snd_cwnd) {
1100 /* Network is feed fully. */
1101 tp->snd_cwnd_used = 0;
1102 tp->snd_cwnd_stamp = tcp_time_stamp;
1103 } else {
1104 /* Network starves. */
1105 if (tp->packets_out > tp->snd_cwnd_used)
1106 tp->snd_cwnd_used = tp->packets_out;
1108 if (sysctl_tcp_slow_start_after_idle &&
1109 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1110 tcp_cwnd_application_limited(sk);
1114 /* Returns the portion of skb which can be sent right away without
1115 * introducing MSS oddities to segment boundaries. In rare cases where
1116 * mss_now != mss_cache, we will request caller to create a small skb
1117 * per input skb which could be mostly avoided here (if desired).
1119 * We explicitly want to create a request for splitting write queue tail
1120 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1121 * thus all the complexity (cwnd_len is always MSS multiple which we
1122 * return whenever allowed by the other factors). Basically we need the
1123 * modulo only when the receiver window alone is the limiting factor or
1124 * when we would be allowed to send the split-due-to-Nagle skb fully.
1126 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1127 unsigned int mss_now, unsigned int cwnd)
1129 struct tcp_sock *tp = tcp_sk(sk);
1130 u32 needed, window, cwnd_len;
1132 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1133 cwnd_len = mss_now * cwnd;
1135 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1136 return cwnd_len;
1138 needed = min(skb->len, window);
1140 if (cwnd_len <= needed)
1141 return cwnd_len;
1143 return needed - needed % mss_now;
1146 /* Can at least one segment of SKB be sent right now, according to the
1147 * congestion window rules? If so, return how many segments are allowed.
1149 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1150 struct sk_buff *skb)
1152 u32 in_flight, cwnd;
1154 /* Don't be strict about the congestion window for the final FIN. */
1155 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1156 tcp_skb_pcount(skb) == 1)
1157 return 1;
1159 in_flight = tcp_packets_in_flight(tp);
1160 cwnd = tp->snd_cwnd;
1161 if (in_flight < cwnd)
1162 return (cwnd - in_flight);
1164 return 0;
1167 /* Intialize TSO state of a skb.
1168 * This must be invoked the first time we consider transmitting
1169 * SKB onto the wire.
1171 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1172 unsigned int mss_now)
1174 int tso_segs = tcp_skb_pcount(skb);
1176 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1177 tcp_set_skb_tso_segs(sk, skb, mss_now);
1178 tso_segs = tcp_skb_pcount(skb);
1180 return tso_segs;
1183 /* Minshall's variant of the Nagle send check. */
1184 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1186 return after(tp->snd_sml, tp->snd_una) &&
1187 !after(tp->snd_sml, tp->snd_nxt);
1190 /* Return 0, if packet can be sent now without violation Nagle's rules:
1191 * 1. It is full sized.
1192 * 2. Or it contains FIN. (already checked by caller)
1193 * 3. Or TCP_NODELAY was set.
1194 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1195 * With Minshall's modification: all sent small packets are ACKed.
1197 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1198 const struct sk_buff *skb,
1199 unsigned mss_now, int nonagle)
1201 return (skb->len < mss_now &&
1202 ((nonagle & TCP_NAGLE_CORK) ||
1203 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1206 /* Return non-zero if the Nagle test allows this packet to be
1207 * sent now.
1209 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1210 unsigned int cur_mss, int nonagle)
1212 /* Nagle rule does not apply to frames, which sit in the middle of the
1213 * write_queue (they have no chances to get new data).
1215 * This is implemented in the callers, where they modify the 'nonagle'
1216 * argument based upon the location of SKB in the send queue.
1218 if (nonagle & TCP_NAGLE_PUSH)
1219 return 1;
1221 /* Don't use the nagle rule for urgent data (or for the final FIN).
1222 * Nagle can be ignored during F-RTO too (see RFC4138).
1224 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1225 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1226 return 1;
1228 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1229 return 1;
1231 return 0;
1234 /* Does at least the first segment of SKB fit into the send window? */
1235 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1236 unsigned int cur_mss)
1238 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1240 if (skb->len > cur_mss)
1241 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1243 return !after(end_seq, tcp_wnd_end(tp));
1246 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1247 * should be put on the wire right now. If so, it returns the number of
1248 * packets allowed by the congestion window.
1250 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1251 unsigned int cur_mss, int nonagle)
1253 struct tcp_sock *tp = tcp_sk(sk);
1254 unsigned int cwnd_quota;
1256 tcp_init_tso_segs(sk, skb, cur_mss);
1258 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1259 return 0;
1261 cwnd_quota = tcp_cwnd_test(tp, skb);
1262 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1263 cwnd_quota = 0;
1265 return cwnd_quota;
1268 /* Test if sending is allowed right now. */
1269 int tcp_may_send_now(struct sock *sk)
1271 struct tcp_sock *tp = tcp_sk(sk);
1272 struct sk_buff *skb = tcp_send_head(sk);
1274 return (skb &&
1275 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1276 (tcp_skb_is_last(sk, skb) ?
1277 tp->nonagle : TCP_NAGLE_PUSH)));
1280 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1281 * which is put after SKB on the list. It is very much like
1282 * tcp_fragment() except that it may make several kinds of assumptions
1283 * in order to speed up the splitting operation. In particular, we
1284 * know that all the data is in scatter-gather pages, and that the
1285 * packet has never been sent out before (and thus is not cloned).
1287 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1288 unsigned int mss_now)
1290 struct sk_buff *buff;
1291 int nlen = skb->len - len;
1292 u8 flags;
1294 /* All of a TSO frame must be composed of paged data. */
1295 if (skb->len != skb->data_len)
1296 return tcp_fragment(sk, skb, len, mss_now);
1298 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1299 if (unlikely(buff == NULL))
1300 return -ENOMEM;
1302 sk->sk_wmem_queued += buff->truesize;
1303 sk_mem_charge(sk, buff->truesize);
1304 buff->truesize += nlen;
1305 skb->truesize -= nlen;
1307 /* Correct the sequence numbers. */
1308 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1309 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1310 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1312 /* PSH and FIN should only be set in the second packet. */
1313 flags = TCP_SKB_CB(skb)->flags;
1314 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1315 TCP_SKB_CB(buff)->flags = flags;
1317 /* This packet was never sent out yet, so no SACK bits. */
1318 TCP_SKB_CB(buff)->sacked = 0;
1320 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1321 skb_split(skb, buff, len);
1323 /* Fix up tso_factor for both original and new SKB. */
1324 tcp_set_skb_tso_segs(sk, skb, mss_now);
1325 tcp_set_skb_tso_segs(sk, buff, mss_now);
1327 /* Link BUFF into the send queue. */
1328 skb_header_release(buff);
1329 tcp_insert_write_queue_after(skb, buff, sk);
1331 return 0;
1334 /* Try to defer sending, if possible, in order to minimize the amount
1335 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1337 * This algorithm is from John Heffner.
1339 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1341 struct tcp_sock *tp = tcp_sk(sk);
1342 const struct inet_connection_sock *icsk = inet_csk(sk);
1343 u32 send_win, cong_win, limit, in_flight;
1345 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1346 goto send_now;
1348 if (icsk->icsk_ca_state != TCP_CA_Open)
1349 goto send_now;
1351 /* Defer for less than two clock ticks. */
1352 if (tp->tso_deferred &&
1353 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1354 goto send_now;
1356 in_flight = tcp_packets_in_flight(tp);
1358 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1360 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1362 /* From in_flight test above, we know that cwnd > in_flight. */
1363 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1365 limit = min(send_win, cong_win);
1367 /* If a full-sized TSO skb can be sent, do it. */
1368 if (limit >= sk->sk_gso_max_size)
1369 goto send_now;
1371 /* Middle in queue won't get any more data, full sendable already? */
1372 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1373 goto send_now;
1375 if (sysctl_tcp_tso_win_divisor) {
1376 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1378 /* If at least some fraction of a window is available,
1379 * just use it.
1381 chunk /= sysctl_tcp_tso_win_divisor;
1382 if (limit >= chunk)
1383 goto send_now;
1384 } else {
1385 /* Different approach, try not to defer past a single
1386 * ACK. Receiver should ACK every other full sized
1387 * frame, so if we have space for more than 3 frames
1388 * then send now.
1390 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1391 goto send_now;
1394 /* Ok, it looks like it is advisable to defer. */
1395 tp->tso_deferred = 1 | (jiffies << 1);
1397 return 1;
1399 send_now:
1400 tp->tso_deferred = 0;
1401 return 0;
1404 /* Create a new MTU probe if we are ready.
1405 * MTU probe is regularly attempting to increase the path MTU by
1406 * deliberately sending larger packets. This discovers routing
1407 * changes resulting in larger path MTUs.
1409 * Returns 0 if we should wait to probe (no cwnd available),
1410 * 1 if a probe was sent,
1411 * -1 otherwise
1413 static int tcp_mtu_probe(struct sock *sk)
1415 struct tcp_sock *tp = tcp_sk(sk);
1416 struct inet_connection_sock *icsk = inet_csk(sk);
1417 struct sk_buff *skb, *nskb, *next;
1418 int len;
1419 int probe_size;
1420 int size_needed;
1421 int copy;
1422 int mss_now;
1424 /* Not currently probing/verifying,
1425 * not in recovery,
1426 * have enough cwnd, and
1427 * not SACKing (the variable headers throw things off) */
1428 if (!icsk->icsk_mtup.enabled ||
1429 icsk->icsk_mtup.probe_size ||
1430 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1431 tp->snd_cwnd < 11 ||
1432 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1433 return -1;
1435 /* Very simple search strategy: just double the MSS. */
1436 mss_now = tcp_current_mss(sk);
1437 probe_size = 2 * tp->mss_cache;
1438 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1439 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1440 /* TODO: set timer for probe_converge_event */
1441 return -1;
1444 /* Have enough data in the send queue to probe? */
1445 if (tp->write_seq - tp->snd_nxt < size_needed)
1446 return -1;
1448 if (tp->snd_wnd < size_needed)
1449 return -1;
1450 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1451 return 0;
1453 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1454 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1455 if (!tcp_packets_in_flight(tp))
1456 return -1;
1457 else
1458 return 0;
1461 /* We're allowed to probe. Build it now. */
1462 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1463 return -1;
1464 sk->sk_wmem_queued += nskb->truesize;
1465 sk_mem_charge(sk, nskb->truesize);
1467 skb = tcp_send_head(sk);
1469 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1470 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1471 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1472 TCP_SKB_CB(nskb)->sacked = 0;
1473 nskb->csum = 0;
1474 nskb->ip_summed = skb->ip_summed;
1476 tcp_insert_write_queue_before(nskb, skb, sk);
1478 len = 0;
1479 tcp_for_write_queue_from_safe(skb, next, sk) {
1480 copy = min_t(int, skb->len, probe_size - len);
1481 if (nskb->ip_summed)
1482 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1483 else
1484 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1485 skb_put(nskb, copy),
1486 copy, nskb->csum);
1488 if (skb->len <= copy) {
1489 /* We've eaten all the data from this skb.
1490 * Throw it away. */
1491 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1492 tcp_unlink_write_queue(skb, sk);
1493 sk_wmem_free_skb(sk, skb);
1494 } else {
1495 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1496 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1497 if (!skb_shinfo(skb)->nr_frags) {
1498 skb_pull(skb, copy);
1499 if (skb->ip_summed != CHECKSUM_PARTIAL)
1500 skb->csum = csum_partial(skb->data,
1501 skb->len, 0);
1502 } else {
1503 __pskb_trim_head(skb, copy);
1504 tcp_set_skb_tso_segs(sk, skb, mss_now);
1506 TCP_SKB_CB(skb)->seq += copy;
1509 len += copy;
1511 if (len >= probe_size)
1512 break;
1514 tcp_init_tso_segs(sk, nskb, nskb->len);
1516 /* We're ready to send. If this fails, the probe will
1517 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1518 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1519 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1520 /* Decrement cwnd here because we are sending
1521 * effectively two packets. */
1522 tp->snd_cwnd--;
1523 tcp_event_new_data_sent(sk, nskb);
1525 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1526 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1527 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1529 return 1;
1532 return -1;
1535 /* This routine writes packets to the network. It advances the
1536 * send_head. This happens as incoming acks open up the remote
1537 * window for us.
1539 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1540 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1541 * account rare use of URG, this is not a big flaw.
1543 * Returns 1, if no segments are in flight and we have queued segments, but
1544 * cannot send anything now because of SWS or another problem.
1546 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1547 int push_one, gfp_t gfp)
1549 struct tcp_sock *tp = tcp_sk(sk);
1550 struct sk_buff *skb;
1551 unsigned int tso_segs, sent_pkts;
1552 int cwnd_quota;
1553 int result;
1555 sent_pkts = 0;
1557 if (!push_one) {
1558 /* Do MTU probing. */
1559 result = tcp_mtu_probe(sk);
1560 if (!result) {
1561 return 0;
1562 } else if (result > 0) {
1563 sent_pkts = 1;
1567 while ((skb = tcp_send_head(sk))) {
1568 unsigned int limit;
1570 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1571 BUG_ON(!tso_segs);
1573 cwnd_quota = tcp_cwnd_test(tp, skb);
1574 if (!cwnd_quota)
1575 break;
1577 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1578 break;
1580 if (tso_segs == 1) {
1581 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1582 (tcp_skb_is_last(sk, skb) ?
1583 nonagle : TCP_NAGLE_PUSH))))
1584 break;
1585 } else {
1586 if (!push_one && tcp_tso_should_defer(sk, skb))
1587 break;
1590 limit = mss_now;
1591 if (tso_segs > 1 && !tcp_urg_mode(tp))
1592 limit = tcp_mss_split_point(sk, skb, mss_now,
1593 cwnd_quota);
1595 if (skb->len > limit &&
1596 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1597 break;
1599 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1601 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1602 break;
1604 /* Advance the send_head. This one is sent out.
1605 * This call will increment packets_out.
1607 tcp_event_new_data_sent(sk, skb);
1609 tcp_minshall_update(tp, mss_now, skb);
1610 sent_pkts++;
1612 if (push_one)
1613 break;
1616 if (likely(sent_pkts)) {
1617 tcp_cwnd_validate(sk);
1618 return 0;
1620 return !tp->packets_out && tcp_send_head(sk);
1623 /* Push out any pending frames which were held back due to
1624 * TCP_CORK or attempt at coalescing tiny packets.
1625 * The socket must be locked by the caller.
1627 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1628 int nonagle)
1630 struct sk_buff *skb = tcp_send_head(sk);
1632 if (!skb)
1633 return;
1635 /* If we are closed, the bytes will have to remain here.
1636 * In time closedown will finish, we empty the write queue and
1637 * all will be happy.
1639 if (unlikely(sk->sk_state == TCP_CLOSE))
1640 return;
1642 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1643 tcp_check_probe_timer(sk);
1646 /* Send _single_ skb sitting at the send head. This function requires
1647 * true push pending frames to setup probe timer etc.
1649 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1651 struct sk_buff *skb = tcp_send_head(sk);
1653 BUG_ON(!skb || skb->len < mss_now);
1655 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1658 /* This function returns the amount that we can raise the
1659 * usable window based on the following constraints
1661 * 1. The window can never be shrunk once it is offered (RFC 793)
1662 * 2. We limit memory per socket
1664 * RFC 1122:
1665 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1666 * RECV.NEXT + RCV.WIN fixed until:
1667 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1669 * i.e. don't raise the right edge of the window until you can raise
1670 * it at least MSS bytes.
1672 * Unfortunately, the recommended algorithm breaks header prediction,
1673 * since header prediction assumes th->window stays fixed.
1675 * Strictly speaking, keeping th->window fixed violates the receiver
1676 * side SWS prevention criteria. The problem is that under this rule
1677 * a stream of single byte packets will cause the right side of the
1678 * window to always advance by a single byte.
1680 * Of course, if the sender implements sender side SWS prevention
1681 * then this will not be a problem.
1683 * BSD seems to make the following compromise:
1685 * If the free space is less than the 1/4 of the maximum
1686 * space available and the free space is less than 1/2 mss,
1687 * then set the window to 0.
1688 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1689 * Otherwise, just prevent the window from shrinking
1690 * and from being larger than the largest representable value.
1692 * This prevents incremental opening of the window in the regime
1693 * where TCP is limited by the speed of the reader side taking
1694 * data out of the TCP receive queue. It does nothing about
1695 * those cases where the window is constrained on the sender side
1696 * because the pipeline is full.
1698 * BSD also seems to "accidentally" limit itself to windows that are a
1699 * multiple of MSS, at least until the free space gets quite small.
1700 * This would appear to be a side effect of the mbuf implementation.
1701 * Combining these two algorithms results in the observed behavior
1702 * of having a fixed window size at almost all times.
1704 * Below we obtain similar behavior by forcing the offered window to
1705 * a multiple of the mss when it is feasible to do so.
1707 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1708 * Regular options like TIMESTAMP are taken into account.
1710 u32 __tcp_select_window(struct sock *sk)
1712 struct inet_connection_sock *icsk = inet_csk(sk);
1713 struct tcp_sock *tp = tcp_sk(sk);
1714 /* MSS for the peer's data. Previous versions used mss_clamp
1715 * here. I don't know if the value based on our guesses
1716 * of peer's MSS is better for the performance. It's more correct
1717 * but may be worse for the performance because of rcv_mss
1718 * fluctuations. --SAW 1998/11/1
1720 int mss = icsk->icsk_ack.rcv_mss;
1721 int free_space = tcp_space(sk);
1722 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1723 int window;
1725 if (mss > full_space)
1726 mss = full_space;
1728 if (free_space < (full_space >> 1)) {
1729 icsk->icsk_ack.quick = 0;
1731 if (tcp_memory_pressure)
1732 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1733 4U * tp->advmss);
1735 if (free_space < mss)
1736 return 0;
1739 if (free_space > tp->rcv_ssthresh)
1740 free_space = tp->rcv_ssthresh;
1742 /* Don't do rounding if we are using window scaling, since the
1743 * scaled window will not line up with the MSS boundary anyway.
1745 window = tp->rcv_wnd;
1746 if (tp->rx_opt.rcv_wscale) {
1747 window = free_space;
1749 /* Advertise enough space so that it won't get scaled away.
1750 * Import case: prevent zero window announcement if
1751 * 1<<rcv_wscale > mss.
1753 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1754 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1755 << tp->rx_opt.rcv_wscale);
1756 } else {
1757 /* Get the largest window that is a nice multiple of mss.
1758 * Window clamp already applied above.
1759 * If our current window offering is within 1 mss of the
1760 * free space we just keep it. This prevents the divide
1761 * and multiply from happening most of the time.
1762 * We also don't do any window rounding when the free space
1763 * is too small.
1765 if (window <= free_space - mss || window > free_space)
1766 window = (free_space / mss) * mss;
1767 else if (mss == full_space &&
1768 free_space > window + (full_space >> 1))
1769 window = free_space;
1772 return window;
1775 /* Collapses two adjacent SKB's during retransmission. */
1776 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1778 struct tcp_sock *tp = tcp_sk(sk);
1779 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1780 int skb_size, next_skb_size;
1782 skb_size = skb->len;
1783 next_skb_size = next_skb->len;
1785 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1787 tcp_highest_sack_combine(sk, next_skb, skb);
1789 tcp_unlink_write_queue(next_skb, sk);
1791 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1792 next_skb_size);
1794 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1795 skb->ip_summed = CHECKSUM_PARTIAL;
1797 if (skb->ip_summed != CHECKSUM_PARTIAL)
1798 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1800 /* Update sequence range on original skb. */
1801 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1803 /* Merge over control information. This moves PSH/FIN etc. over */
1804 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
1806 /* All done, get rid of second SKB and account for it so
1807 * packet counting does not break.
1809 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1811 /* changed transmit queue under us so clear hints */
1812 tcp_clear_retrans_hints_partial(tp);
1813 if (next_skb == tp->retransmit_skb_hint)
1814 tp->retransmit_skb_hint = skb;
1816 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
1818 sk_wmem_free_skb(sk, next_skb);
1821 /* Check if coalescing SKBs is legal. */
1822 static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
1824 if (tcp_skb_pcount(skb) > 1)
1825 return 0;
1826 /* TODO: SACK collapsing could be used to remove this condition */
1827 if (skb_shinfo(skb)->nr_frags != 0)
1828 return 0;
1829 if (skb_cloned(skb))
1830 return 0;
1831 if (skb == tcp_send_head(sk))
1832 return 0;
1833 /* Some heurestics for collapsing over SACK'd could be invented */
1834 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1835 return 0;
1837 return 1;
1840 /* Collapse packets in the retransmit queue to make to create
1841 * less packets on the wire. This is only done on retransmission.
1843 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
1844 int space)
1846 struct tcp_sock *tp = tcp_sk(sk);
1847 struct sk_buff *skb = to, *tmp;
1848 int first = 1;
1850 if (!sysctl_tcp_retrans_collapse)
1851 return;
1852 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)
1853 return;
1855 tcp_for_write_queue_from_safe(skb, tmp, sk) {
1856 if (!tcp_can_collapse(sk, skb))
1857 break;
1859 space -= skb->len;
1861 if (first) {
1862 first = 0;
1863 continue;
1866 if (space < 0)
1867 break;
1868 /* Punt if not enough space exists in the first SKB for
1869 * the data in the second
1871 if (skb->len > skb_tailroom(to))
1872 break;
1874 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
1875 break;
1877 tcp_collapse_retrans(sk, to);
1881 /* This retransmits one SKB. Policy decisions and retransmit queue
1882 * state updates are done by the caller. Returns non-zero if an
1883 * error occurred which prevented the send.
1885 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1887 struct tcp_sock *tp = tcp_sk(sk);
1888 struct inet_connection_sock *icsk = inet_csk(sk);
1889 unsigned int cur_mss;
1890 int err;
1892 /* Inconslusive MTU probe */
1893 if (icsk->icsk_mtup.probe_size) {
1894 icsk->icsk_mtup.probe_size = 0;
1897 /* Do not sent more than we queued. 1/4 is reserved for possible
1898 * copying overhead: fragmentation, tunneling, mangling etc.
1900 if (atomic_read(&sk->sk_wmem_alloc) >
1901 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1902 return -EAGAIN;
1904 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1905 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1906 BUG();
1907 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1908 return -ENOMEM;
1911 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1912 return -EHOSTUNREACH; /* Routing failure or similar. */
1914 cur_mss = tcp_current_mss(sk);
1916 /* If receiver has shrunk his window, and skb is out of
1917 * new window, do not retransmit it. The exception is the
1918 * case, when window is shrunk to zero. In this case
1919 * our retransmit serves as a zero window probe.
1921 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1922 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1923 return -EAGAIN;
1925 if (skb->len > cur_mss) {
1926 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1927 return -ENOMEM; /* We'll try again later. */
1928 } else {
1929 int oldpcount = tcp_skb_pcount(skb);
1931 if (unlikely(oldpcount > 1)) {
1932 tcp_init_tso_segs(sk, skb, cur_mss);
1933 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
1937 tcp_retrans_try_collapse(sk, skb, cur_mss);
1939 /* Some Solaris stacks overoptimize and ignore the FIN on a
1940 * retransmit when old data is attached. So strip it off
1941 * since it is cheap to do so and saves bytes on the network.
1943 if (skb->len > 0 &&
1944 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1945 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1946 if (!pskb_trim(skb, 0)) {
1947 /* Reuse, even though it does some unnecessary work */
1948 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1949 TCP_SKB_CB(skb)->flags);
1950 skb->ip_summed = CHECKSUM_NONE;
1954 /* Make a copy, if the first transmission SKB clone we made
1955 * is still in somebody's hands, else make a clone.
1957 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1959 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1961 if (err == 0) {
1962 /* Update global TCP statistics. */
1963 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
1965 tp->total_retrans++;
1967 #if FASTRETRANS_DEBUG > 0
1968 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1969 if (net_ratelimit())
1970 printk(KERN_DEBUG "retrans_out leaked.\n");
1972 #endif
1973 if (!tp->retrans_out)
1974 tp->lost_retrans_low = tp->snd_nxt;
1975 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1976 tp->retrans_out += tcp_skb_pcount(skb);
1978 /* Save stamp of the first retransmit. */
1979 if (!tp->retrans_stamp)
1980 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1982 tp->undo_retrans++;
1984 /* snd_nxt is stored to detect loss of retransmitted segment,
1985 * see tcp_input.c tcp_sacktag_write_queue().
1987 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1989 return err;
1992 /* Check if we forward retransmits are possible in the current
1993 * window/congestion state.
1995 static int tcp_can_forward_retransmit(struct sock *sk)
1997 const struct inet_connection_sock *icsk = inet_csk(sk);
1998 struct tcp_sock *tp = tcp_sk(sk);
2000 /* Forward retransmissions are possible only during Recovery. */
2001 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2002 return 0;
2004 /* No forward retransmissions in Reno are possible. */
2005 if (tcp_is_reno(tp))
2006 return 0;
2008 /* Yeah, we have to make difficult choice between forward transmission
2009 * and retransmission... Both ways have their merits...
2011 * For now we do not retransmit anything, while we have some new
2012 * segments to send. In the other cases, follow rule 3 for
2013 * NextSeg() specified in RFC3517.
2016 if (tcp_may_send_now(sk))
2017 return 0;
2019 return 1;
2022 /* This gets called after a retransmit timeout, and the initially
2023 * retransmitted data is acknowledged. It tries to continue
2024 * resending the rest of the retransmit queue, until either
2025 * we've sent it all or the congestion window limit is reached.
2026 * If doing SACK, the first ACK which comes back for a timeout
2027 * based retransmit packet might feed us FACK information again.
2028 * If so, we use it to avoid unnecessarily retransmissions.
2030 void tcp_xmit_retransmit_queue(struct sock *sk)
2032 const struct inet_connection_sock *icsk = inet_csk(sk);
2033 struct tcp_sock *tp = tcp_sk(sk);
2034 struct sk_buff *skb;
2035 struct sk_buff *hole = NULL;
2036 u32 last_lost;
2037 int mib_idx;
2038 int fwd_rexmitting = 0;
2040 if (!tp->lost_out)
2041 tp->retransmit_high = tp->snd_una;
2043 if (tp->retransmit_skb_hint) {
2044 skb = tp->retransmit_skb_hint;
2045 last_lost = TCP_SKB_CB(skb)->end_seq;
2046 if (after(last_lost, tp->retransmit_high))
2047 last_lost = tp->retransmit_high;
2048 } else {
2049 skb = tcp_write_queue_head(sk);
2050 last_lost = tp->snd_una;
2053 tcp_for_write_queue_from(skb, sk) {
2054 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2056 if (skb == tcp_send_head(sk))
2057 break;
2058 /* we could do better than to assign each time */
2059 if (hole == NULL)
2060 tp->retransmit_skb_hint = skb;
2062 /* Assume this retransmit will generate
2063 * only one packet for congestion window
2064 * calculation purposes. This works because
2065 * tcp_retransmit_skb() will chop up the
2066 * packet to be MSS sized and all the
2067 * packet counting works out.
2069 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2070 return;
2072 if (fwd_rexmitting) {
2073 begin_fwd:
2074 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2075 break;
2076 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2078 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2079 tp->retransmit_high = last_lost;
2080 if (!tcp_can_forward_retransmit(sk))
2081 break;
2082 /* Backtrack if necessary to non-L'ed skb */
2083 if (hole != NULL) {
2084 skb = hole;
2085 hole = NULL;
2087 fwd_rexmitting = 1;
2088 goto begin_fwd;
2090 } else if (!(sacked & TCPCB_LOST)) {
2091 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2092 hole = skb;
2093 continue;
2095 } else {
2096 last_lost = TCP_SKB_CB(skb)->end_seq;
2097 if (icsk->icsk_ca_state != TCP_CA_Loss)
2098 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2099 else
2100 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2103 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2104 continue;
2106 if (tcp_retransmit_skb(sk, skb))
2107 return;
2108 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2110 if (skb == tcp_write_queue_head(sk))
2111 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2112 inet_csk(sk)->icsk_rto,
2113 TCP_RTO_MAX);
2117 /* Send a fin. The caller locks the socket for us. This cannot be
2118 * allowed to fail queueing a FIN frame under any circumstances.
2120 void tcp_send_fin(struct sock *sk)
2122 struct tcp_sock *tp = tcp_sk(sk);
2123 struct sk_buff *skb = tcp_write_queue_tail(sk);
2124 int mss_now;
2126 /* Optimization, tack on the FIN if we have a queue of
2127 * unsent frames. But be careful about outgoing SACKS
2128 * and IP options.
2130 mss_now = tcp_current_mss(sk);
2132 if (tcp_send_head(sk) != NULL) {
2133 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2134 TCP_SKB_CB(skb)->end_seq++;
2135 tp->write_seq++;
2136 } else {
2137 /* Socket is locked, keep trying until memory is available. */
2138 for (;;) {
2139 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2140 sk->sk_allocation);
2141 if (skb)
2142 break;
2143 yield();
2146 /* Reserve space for headers and prepare control bits. */
2147 skb_reserve(skb, MAX_TCP_HEADER);
2148 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2149 tcp_init_nondata_skb(skb, tp->write_seq,
2150 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2151 tcp_queue_skb(sk, skb);
2153 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2156 /* We get here when a process closes a file descriptor (either due to
2157 * an explicit close() or as a byproduct of exit()'ing) and there
2158 * was unread data in the receive queue. This behavior is recommended
2159 * by RFC 2525, section 2.17. -DaveM
2161 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2163 struct sk_buff *skb;
2165 /* NOTE: No TCP options attached and we never retransmit this. */
2166 skb = alloc_skb(MAX_TCP_HEADER, priority);
2167 if (!skb) {
2168 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2169 return;
2172 /* Reserve space for headers and prepare control bits. */
2173 skb_reserve(skb, MAX_TCP_HEADER);
2174 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2175 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2176 /* Send it off. */
2177 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2178 if (tcp_transmit_skb(sk, skb, 0, priority))
2179 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2181 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2184 /* Send a crossed SYN-ACK during socket establishment.
2185 * WARNING: This routine must only be called when we have already sent
2186 * a SYN packet that crossed the incoming SYN that caused this routine
2187 * to get called. If this assumption fails then the initial rcv_wnd
2188 * and rcv_wscale values will not be correct.
2190 int tcp_send_synack(struct sock *sk)
2192 struct sk_buff *skb;
2194 skb = tcp_write_queue_head(sk);
2195 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2196 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2197 return -EFAULT;
2199 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2200 if (skb_cloned(skb)) {
2201 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2202 if (nskb == NULL)
2203 return -ENOMEM;
2204 tcp_unlink_write_queue(skb, sk);
2205 skb_header_release(nskb);
2206 __tcp_add_write_queue_head(sk, nskb);
2207 sk_wmem_free_skb(sk, skb);
2208 sk->sk_wmem_queued += nskb->truesize;
2209 sk_mem_charge(sk, nskb->truesize);
2210 skb = nskb;
2213 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2214 TCP_ECN_send_synack(tcp_sk(sk), skb);
2216 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2217 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2220 /* Prepare a SYN-ACK. */
2221 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2222 struct request_sock *req)
2224 struct inet_request_sock *ireq = inet_rsk(req);
2225 struct tcp_sock *tp = tcp_sk(sk);
2226 struct tcphdr *th;
2227 int tcp_header_size;
2228 struct tcp_out_options opts;
2229 struct sk_buff *skb;
2230 struct tcp_md5sig_key *md5;
2231 __u8 *md5_hash_location;
2232 int mss;
2234 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2235 if (skb == NULL)
2236 return NULL;
2238 /* Reserve space for headers. */
2239 skb_reserve(skb, MAX_TCP_HEADER);
2241 skb_dst_set(skb, dst_clone(dst));
2243 mss = dst_metric(dst, RTAX_ADVMSS);
2244 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2245 mss = tp->rx_opt.user_mss;
2247 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2248 __u8 rcv_wscale;
2249 /* Set this up on the first call only */
2250 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2251 /* tcp_full_space because it is guaranteed to be the first packet */
2252 tcp_select_initial_window(tcp_full_space(sk),
2253 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2254 &req->rcv_wnd,
2255 &req->window_clamp,
2256 ireq->wscale_ok,
2257 &rcv_wscale);
2258 ireq->rcv_wscale = rcv_wscale;
2261 memset(&opts, 0, sizeof(opts));
2262 #ifdef CONFIG_SYN_COOKIES
2263 if (unlikely(req->cookie_ts))
2264 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2265 else
2266 #endif
2267 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2268 tcp_header_size = tcp_synack_options(sk, req, mss,
2269 skb, &opts, &md5) +
2270 sizeof(struct tcphdr);
2272 skb_push(skb, tcp_header_size);
2273 skb_reset_transport_header(skb);
2275 th = tcp_hdr(skb);
2276 memset(th, 0, sizeof(struct tcphdr));
2277 th->syn = 1;
2278 th->ack = 1;
2279 TCP_ECN_make_synack(req, th);
2280 th->source = ireq->loc_port;
2281 th->dest = ireq->rmt_port;
2282 /* Setting of flags are superfluous here for callers (and ECE is
2283 * not even correctly set)
2285 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2286 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
2287 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2288 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2290 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2291 th->window = htons(min(req->rcv_wnd, 65535U));
2292 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
2293 th->doff = (tcp_header_size >> 2);
2294 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
2296 #ifdef CONFIG_TCP_MD5SIG
2297 /* Okay, we have all we need - do the md5 hash if needed */
2298 if (md5) {
2299 tcp_rsk(req)->af_specific->calc_md5_hash(md5_hash_location,
2300 md5, NULL, req, skb);
2302 #endif
2304 return skb;
2307 /* Do all connect socket setups that can be done AF independent. */
2308 static void tcp_connect_init(struct sock *sk)
2310 struct dst_entry *dst = __sk_dst_get(sk);
2311 struct tcp_sock *tp = tcp_sk(sk);
2312 __u8 rcv_wscale;
2314 /* We'll fix this up when we get a response from the other end.
2315 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2317 tp->tcp_header_len = sizeof(struct tcphdr) +
2318 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2320 #ifdef CONFIG_TCP_MD5SIG
2321 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2322 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2323 #endif
2325 /* If user gave his TCP_MAXSEG, record it to clamp */
2326 if (tp->rx_opt.user_mss)
2327 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2328 tp->max_window = 0;
2329 tcp_mtup_init(sk);
2330 tcp_sync_mss(sk, dst_mtu(dst));
2332 if (!tp->window_clamp)
2333 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2334 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2335 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2336 tp->advmss = tp->rx_opt.user_mss;
2338 tcp_initialize_rcv_mss(sk);
2340 tcp_select_initial_window(tcp_full_space(sk),
2341 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2342 &tp->rcv_wnd,
2343 &tp->window_clamp,
2344 sysctl_tcp_window_scaling,
2345 &rcv_wscale);
2347 tp->rx_opt.rcv_wscale = rcv_wscale;
2348 tp->rcv_ssthresh = tp->rcv_wnd;
2350 sk->sk_err = 0;
2351 sock_reset_flag(sk, SOCK_DONE);
2352 tp->snd_wnd = 0;
2353 tcp_init_wl(tp, 0);
2354 tp->snd_una = tp->write_seq;
2355 tp->snd_sml = tp->write_seq;
2356 tp->snd_up = tp->write_seq;
2357 tp->rcv_nxt = 0;
2358 tp->rcv_wup = 0;
2359 tp->copied_seq = 0;
2361 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2362 inet_csk(sk)->icsk_retransmits = 0;
2363 tcp_clear_retrans(tp);
2366 /* Build a SYN and send it off. */
2367 int tcp_connect(struct sock *sk)
2369 struct tcp_sock *tp = tcp_sk(sk);
2370 struct sk_buff *buff;
2372 tcp_connect_init(sk);
2374 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2375 if (unlikely(buff == NULL))
2376 return -ENOBUFS;
2378 /* Reserve space for headers. */
2379 skb_reserve(buff, MAX_TCP_HEADER);
2381 tp->snd_nxt = tp->write_seq;
2382 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2383 TCP_ECN_send_syn(sk, buff);
2385 /* Send it off. */
2386 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2387 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2388 skb_header_release(buff);
2389 __tcp_add_write_queue_tail(sk, buff);
2390 sk->sk_wmem_queued += buff->truesize;
2391 sk_mem_charge(sk, buff->truesize);
2392 tp->packets_out += tcp_skb_pcount(buff);
2393 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2395 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2396 * in order to make this packet get counted in tcpOutSegs.
2398 tp->snd_nxt = tp->write_seq;
2399 tp->pushed_seq = tp->write_seq;
2400 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2402 /* Timer for repeating the SYN until an answer. */
2403 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2404 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2405 return 0;
2408 /* Send out a delayed ack, the caller does the policy checking
2409 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2410 * for details.
2412 void tcp_send_delayed_ack(struct sock *sk)
2414 struct inet_connection_sock *icsk = inet_csk(sk);
2415 int ato = icsk->icsk_ack.ato;
2416 unsigned long timeout;
2418 if (ato > TCP_DELACK_MIN) {
2419 const struct tcp_sock *tp = tcp_sk(sk);
2420 int max_ato = HZ / 2;
2422 if (icsk->icsk_ack.pingpong ||
2423 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2424 max_ato = TCP_DELACK_MAX;
2426 /* Slow path, intersegment interval is "high". */
2428 /* If some rtt estimate is known, use it to bound delayed ack.
2429 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2430 * directly.
2432 if (tp->srtt) {
2433 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2435 if (rtt < max_ato)
2436 max_ato = rtt;
2439 ato = min(ato, max_ato);
2442 /* Stay within the limit we were given */
2443 timeout = jiffies + ato;
2445 /* Use new timeout only if there wasn't a older one earlier. */
2446 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2447 /* If delack timer was blocked or is about to expire,
2448 * send ACK now.
2450 if (icsk->icsk_ack.blocked ||
2451 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2452 tcp_send_ack(sk);
2453 return;
2456 if (!time_before(timeout, icsk->icsk_ack.timeout))
2457 timeout = icsk->icsk_ack.timeout;
2459 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2460 icsk->icsk_ack.timeout = timeout;
2461 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2464 /* This routine sends an ack and also updates the window. */
2465 void tcp_send_ack(struct sock *sk)
2467 struct sk_buff *buff;
2469 /* If we have been reset, we may not send again. */
2470 if (sk->sk_state == TCP_CLOSE)
2471 return;
2473 /* We are not putting this on the write queue, so
2474 * tcp_transmit_skb() will set the ownership to this
2475 * sock.
2477 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2478 if (buff == NULL) {
2479 inet_csk_schedule_ack(sk);
2480 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2481 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2482 TCP_DELACK_MAX, TCP_RTO_MAX);
2483 return;
2486 /* Reserve space for headers and prepare control bits. */
2487 skb_reserve(buff, MAX_TCP_HEADER);
2488 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
2490 /* Send it off, this clears delayed acks for us. */
2491 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2492 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2495 /* This routine sends a packet with an out of date sequence
2496 * number. It assumes the other end will try to ack it.
2498 * Question: what should we make while urgent mode?
2499 * 4.4BSD forces sending single byte of data. We cannot send
2500 * out of window data, because we have SND.NXT==SND.MAX...
2502 * Current solution: to send TWO zero-length segments in urgent mode:
2503 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2504 * out-of-date with SND.UNA-1 to probe window.
2506 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2508 struct tcp_sock *tp = tcp_sk(sk);
2509 struct sk_buff *skb;
2511 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2512 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2513 if (skb == NULL)
2514 return -1;
2516 /* Reserve space for headers and set control bits. */
2517 skb_reserve(skb, MAX_TCP_HEADER);
2518 /* Use a previous sequence. This should cause the other
2519 * end to send an ack. Don't queue or clone SKB, just
2520 * send it.
2522 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
2523 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2524 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2527 /* Initiate keepalive or window probe from timer. */
2528 int tcp_write_wakeup(struct sock *sk)
2530 struct tcp_sock *tp = tcp_sk(sk);
2531 struct sk_buff *skb;
2533 if (sk->sk_state == TCP_CLOSE)
2534 return -1;
2536 if ((skb = tcp_send_head(sk)) != NULL &&
2537 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2538 int err;
2539 unsigned int mss = tcp_current_mss(sk);
2540 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2542 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2543 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2545 /* We are probing the opening of a window
2546 * but the window size is != 0
2547 * must have been a result SWS avoidance ( sender )
2549 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2550 skb->len > mss) {
2551 seg_size = min(seg_size, mss);
2552 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2553 if (tcp_fragment(sk, skb, seg_size, mss))
2554 return -1;
2555 } else if (!tcp_skb_pcount(skb))
2556 tcp_set_skb_tso_segs(sk, skb, mss);
2558 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2559 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2560 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2561 if (!err)
2562 tcp_event_new_data_sent(sk, skb);
2563 return err;
2564 } else {
2565 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2566 tcp_xmit_probe_skb(sk, 1);
2567 return tcp_xmit_probe_skb(sk, 0);
2571 /* A window probe timeout has occurred. If window is not closed send
2572 * a partial packet else a zero probe.
2574 void tcp_send_probe0(struct sock *sk)
2576 struct inet_connection_sock *icsk = inet_csk(sk);
2577 struct tcp_sock *tp = tcp_sk(sk);
2578 int err;
2580 err = tcp_write_wakeup(sk);
2582 if (tp->packets_out || !tcp_send_head(sk)) {
2583 /* Cancel probe timer, if it is not required. */
2584 icsk->icsk_probes_out = 0;
2585 icsk->icsk_backoff = 0;
2586 return;
2589 if (err <= 0) {
2590 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2591 icsk->icsk_backoff++;
2592 icsk->icsk_probes_out++;
2593 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2594 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2595 TCP_RTO_MAX);
2596 } else {
2597 /* If packet was not sent due to local congestion,
2598 * do not backoff and do not remember icsk_probes_out.
2599 * Let local senders to fight for local resources.
2601 * Use accumulated backoff yet.
2603 if (!icsk->icsk_probes_out)
2604 icsk->icsk_probes_out = 1;
2605 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2606 min(icsk->icsk_rto << icsk->icsk_backoff,
2607 TCP_RESOURCE_PROBE_INTERVAL),
2608 TCP_RTO_MAX);
2612 EXPORT_SYMBOL(tcp_select_initial_window);
2613 EXPORT_SYMBOL(tcp_connect);
2614 EXPORT_SYMBOL(tcp_make_synack);
2615 EXPORT_SYMBOL(tcp_simple_retransmit);
2616 EXPORT_SYMBOL(tcp_sync_mss);
2617 EXPORT_SYMBOL(tcp_mtup_init);